Unbiased Protein Interface Prediction Based on Ligand Diversity Quantification∗ Reyhaneh Esmaielbeiki and Jean-Christophe Nebel Faculty of Science, Engineering and Computing, Kingston University, London Kingston-Upon-Thames, Surrey, KT1 2EE, UK {r.esmaielbeiki,J.Nebel}@kingston.ac.uk Abstract Proteins interact with each other to perform essential functions in cells. Consequently, identific- ation of their binding interfaces can provide key information for drug design. Here, we introduce Weighted Protein Interface Prediction (WePIP), an original framework which predicts protein interfaces from homologous complexes. WePIP takes advantage of a novel weighted score which is not only based on structural neighbours’ information but, unlike current state-of-the-art methods, also takes into consideration the nature of their interaction partners. Experimental validation demonstrates that our weighted schema significantly improves prediction performance. In par- ticular, we have established a major contribution to ligand diversity quantification. Moreover, application of our framework on a standard dataset shows WePIP performance compares favour- ably with other state of the art methods. 1998 ACM Subject Classification J.3 Life and Medical Sciences (Biology and genetics) Keywords and phrases Protein-protein interaction, protein interface prediction, homology mod- eling Digital Object Identifier 10.4230/OASIcs.GCB.2012.119 1 Introduction Protein-protein interaction (PPIs) is essential for the functionality of living cells. Alterations of these interactions affect biochemical processes which may lead to critical diseases such as cancer [31]. Therefore, knowledge about protein interactions and their resulting 3D complexes can provide key information for drug design. A number of experimental techniques are available to identify residues involved in PPIs [31]. Although they provide valuable contribution to PPI knowledge, their cost in terms of time and expense limits their practical use [10]. Consequently, computational methods have been proposed to identify protein interfaces. They can be broadly divided in sequence only and structure based approaches. Sequence based methodologies usually rely on a sliding window which allows calculating specific features associated to each amino acid according to its neighbours [24, 28, 35, 8]. Then, a classifier discriminates between interface and non-interface residues according to residue scores. Those approaches differ mainly in their selection of amino acid properties, such as physico-chemical properties [8, 7], residues distribution [24] or conservation degree [34, 25], machine learning algorithm and scoring functions [38]. When the 3D structure of the query protein (QP) is available, integration of structural information, e.g. residues secondary structure or solvent-accessible surface area, allows ∗ This work was in part supported by grant 6435/B/T02/2011/40 of the Polish National Centre for Science. © Reyhaneh. Esmaielbeiki and Jean-Christophe Nebel; licensed under Creative Commons License ND German Conference on Bioinformatics 2012 (GCB’12). Editors: S. Böcker, F. Hufsky, K. Scheubert, J. Schleicher, S. Schuster; pp. 119–130 OpenAccess Series in Informatics Schloss Dagstuhl – Leibniz-Zentrum für Informatik, Dagstuhl Publishing, Germany 120 Unbiased Protein Interface Prediction better predictions [25, 32]. Three approaches taking advantage of these properties are seen as state of the art [38]. ProMate combines 13 different properties, such as chemical component, geometric properties and information from relevant crystal structures, to generate a quantitative measure [23]. Protein interface residues are then predicted using a clustering process relying on mutual information. Alternatively, Cons-PPISP discriminates between residues using neural networks trained with protein’s surface sequence profiles and solvent accessibility of neighbouring residues [6]. Finally, PINUP addresses the problem using an empirical energy function which is based on a linear combination of side chain energy score, interface propensity and residue conservation [22]. Despite fundamental differences, these three approaches display very similar performance [38]. However, each of them seems to capture different important aspects of residue interactions. As a result, a meta-predictor, Meta-PPISP, combining their scores using a linear regression analysis, is able to outperform each of these individual methods in terms of accuracy [27]. With the increasing number of experimentally determined protein 3D structures, they have become the main source of interface prediction methods. First, structurally homologous proteins tend to display similar interaction sites [1, 9]. Secondly, protein’s binding sites are evolutionarily conserved among structurally similar proteins (or structural neighbours) [37, 18, 19, 4, 36, 5]. Even remote structural neighbours have been shown to display a significant level of interface conservation [37]. Consequently, structure based methods for interface prediction rely on analysing proteins which are structurally similar to the query protein. Initial approaches focused on detecting conserved areas among homologous structural neighbours. Carl et al. use graph representation of surface residues [18, 19] to describe homo- logous binding sites [4]. They then refine their technique by using local structural similarity instead of global similarities of the query protein to detect the structural neighbours [5]. A more general method, PredUs, maps interacting residues from structural neighbours onto QP even if they do not display any homology [36]. Whereas PredUs still dependents on the existence of structural neighbours of QP, PrISE proposes to deal with this limitation by predicting interface residues from local structural similarity only [15]. This is achieved using a repository of structural elements (SE) generated from the Protein Data Bank (PDB) [3]. For each SE of the query protein (consisting of a central residues and it neighbours) a set of similar SEs are extracted from the repository and a weight is assigned to them based on their similarity to QP. The central residues of the query protein’s SE are predicted as interface if a weighted majority of its similar SEs are interface residues. Although more general, PrISE displays comparable performance to PredUs [15]. Those two approaches have significantly improved the ability of predicting interface residues; see Table 3. However, they do not deal satisfactorily with the very heterogeneous nature of the PDB. First, the presence of complex duplicates, or homologs, biases predictions towards specific configurations, which can affect negatively performance. Secondly, confidence in the information provided by the interface of a structural neighbour should depend on its degree of homology with QP. Although PrISE acknowledges both issues, it does not address the first one [15]. PredUs deals with these matters in a binary fashion. Complexes involving structural neighbours are clustered and a 40% similarity cut-off is used to choose the representatives which will inform interface prediction. Here, we address those limitations of structure based methods by quantifying, first, homology between QP and its structural neighbours and,second, ligand diversity between the partners, or ligands, of the structural neighbours. In this study we introduce Weighted Protein Interface Prediction (WePIP) framework, a novel PIP approach based on structural neighbours’ information. Its main contribution R. Esmaielbeiki and J.-C. Nebel 121 Figure 1 Interface prediction framework for single (A) and pair protein queries (B). A) For a single query, if at least one homologous complex exists, interfaces are predicted using WePIP otherwise PredUs is used. B) For a pair of proteins, depending on the existence of homologous complexes, interfaces are predicted by one or a combination of the WePIPc, WePIP and PredUs methods. is the weighted score assigned to each residue of QP, which takes into account not only the degree of homology of structural neighbours, but also the nature of their interacting partners. After description of the methodology and validation of the proposed weighted schema, WePIP is evaluated against state of art protein interface prediction methods using a standard benchmark dataset. 2 Methods 2.1 Interface Prediction Principles When two protein chains form a dimer, they bind through their interaction interfaces. We propose a novel methodology predicting the amino acids which are involved in binding interactions based on the 3D structures of the dimer partners. In this study dimer refers to any two protein chains involved in interaction. Not only does our approach predict the locations of interfaces when both binding partners are known, but it also infers the most likely binding interface of a single protein. Figure 1.A and 1.B describe those interface prediction pipelines for single and pair protein queries respectively. Our methodology relies on discovering interaction patterns from the analysis of the 3D structures of complexes involving homologs of the dimer partners, called ‘homologous complexes’. In this work, proteins are defined as homologous if their sequence similarity is expressed by an E-value ≤ 10−2. First, Blast [2] is used to classify each query partner G C B 2 0 1 2 122 Unbiased Protein Interface Prediction according to the availability of homologous complexes in PDB [3]. When both interaction partners are known and common homologous complexes exist, the WePIPc method exploits